Why it matters: As first-person drone racing (FPV) has become more popular, AI implementations have continued to improve their results compared to human pilots. While there is still a lot of uncharted territory for this research area, it could ultimately have implications for various real-world applications of autonomous drones.
In 2021, researchers from the University of Zurich presented an autonomous drone control system that could overtake human pilots on racetracks. In the two years since, they’ve done it developed a successor they claim has bested three world champion FPV drone racers.
The burgeoning sport challenges contestants to fly a small drone through a series of gates in the correct order as quickly as possible, with the video feed from the drone’s camera linked to the pilot’s goggles. The quick reflexes and high level of skill displayed by experienced racers pushes the limits of drone maneuverability, making them an interesting target for research into autonomous control systems.
Called Swift, the AI’s training involved a neural network and data received from an onboard computer, camera and inertial sensor. Swift set record times on the track during testing and defeated three international world champions, largely because he made far tighter turns than the human pilots. Research on autonomous racing systems is almost as old as drone racing, but recent results from the University of Zurich have taken it to a new level.
Perhaps the most striking factor is that while the human racers trained on the test track for a week, the AI training process only took about an hour on a standard workstation desktop. Two possible advantages for the drone are that it processes information faster than racers’ brains and perceives inertia in a way that humans don’t. However, Swift’s video feed was only 30Hz, while the pilots’ cameras were updated at 120Hz, giving them more visual data.
A major caveat is that Swift has only been tested on an indoor track, while drone races take place in a variety of indoor and outdoor settings. It’s unclear how autonomous systems like Swift would deal with factors like wind or changes in light levels, so there’s certainly room for future research.
The results of this and other experiments could have implications far beyond drone racing. They could help improve the navigation of self-flying drones in real-world environments for purposes like delivery, search and rescue, warfare, and more.